基于符号表示的可度量shapelets提取的时序分类研究

Measurable Shapelets Extraction Based on Symbolic Rrepresentation for Time Series Classification

在时序分类问题中,基于符号表示的shapelets提取方法具有良好的分类精度和分类效率,但对符号进行质量度量的过程,如计算TFIDF分数,耗时较长且计算量大,导致分类效率较低。此外,提取的shapelets候选数量仍然较多,判别力有待提高。针对这些问题,本文提出了一种基于符号表示的可度量shapelets提取方法,该方法包含时间序列数据预处理、确定shapelets候选集和学习shapelets 3个阶段,可以快速得到高质量shapelets。在数据预处理阶段,将时间序列转化为符号聚合近似(SAX)表示以降低原始时间序列的维度。在确定shapelets候选集阶段,利用Bloom过滤器过滤重复的SAX词,并将过滤后的SAX词存储在哈希表中进行质量度量。随后,对SAX词的相似性进行判别,基于相似性和覆盖度等概念确定最终的shapelets候选集。在学习shapelets阶段,采用logistic回归模型学得真正的shapelets用于时序分类。在32个数据集上进行了大量实验,实验结果表明,所提方法的平均分类精度和平均分类效率均排名第二。与现有的基于shapelets的时序分类方法相比,该方法可以在保证精度的同时提高分类效率,并且具有良好的可解释性。

In the time series classification problems,shapelets extraction method based on symbol representation has good classification accuracy and efficiency,but the quality measurement of symbols,such as calculating TFIDF scores,is time-consuming and computatively heavy,leading to low classification efficiency.In addition,there are still a large number of shapelets candidates extracted,and the discriminating power needs to be improved.To solve these problems,this paper proposes a measurable shapelets extraction method based on symbolic representation,which includes three stages:time series data preprocessing,determining shapelets candidate set and learning shapelets,so that high-quality shapelets can be obtained quickly.In the data preprocessing stage,the time series is transformed into a symbolic aggregation approximation(SAX)representation to reduce the dimensions of the original time series.In the stage of determining the candidate set of shapelets,Bloom filters are used to filter repeated SAX words,and the filtered SAX words are stored in the hash table for quality measurement.Then,the similarity of SAX words is discriminated,and the final shapelets candidate set is determined based on the concepts of similarity and coverage.In the learning phase of shapelets,the logistic regression model is used to learn real shapelets for time series classification.In this paper,a large number of experiments are conducted on 32 datasets,and the experimental results show that the average classification accuracy and average classification efficiency of the proposed method rank second on 32 datasets.Compared with the existing time series classification methods based on shapelets,the proposed method can improve the classification efficiency while ensuring the accuracy,and has good interpretability.